Meta-Agent Ensembles: Orchestrating Multi-Agent Collaboration for Diverse CUDA Kernel Optimization

TL;DR: What if we had a team of specialized RL agents, each excelling at different optimization tactics, working together to generate the best CUDA code? The idea is to coordinate multiple, diverse RL agents—some focusing on memory, others on compute, or warp scheduling—using a meta-agent that learns how to combine their strengths for each kernel. An initial experiment could pit this ensemble against single-agent and compiler baselines on the hardest KernelBench tasks.

Research Question: Can a meta-agent that orchestrates a diverse ensemble of specialized RL agents outperform monolithic RL or compiler-based systems in optimizing CUDA kernel generation across heterogeneous workloads?

Hypothesis: A coordinated ensemble of skill-specialized RL agents, guided by a learned meta-agent, will yield superior CUDA kernel performance and generalization compared to single-agent or compiler-based approaches, especially for complex or novel workloads.

Experiment Plan: - Setup: Develop multiple RL agents, each trained to optimize a specific aspect of CUDA kernels (memory access, instruction-level parallelism, occupancy, etc.).

• Meta-Agent: Train a top-level agent that selects which specialist(s) to invoke at each optimization stage.
• Data: Use KernelBench and diverse custom workloads.
• Measurements: Compare speedups, code validity, and generalization across unseen hardware.
• Expected Outcome: The ensemble approach should produce kernels that outperform both single-agent RL and torch.compile, especially on challenging or out-of-distribution tasks.

References:

• 1. Dai, W., Wu, H., Yu, Q., Gao, H., Li, J., Jiang, C., Lou, W., Song, Y., Yu, H., Chen, J., Ma, W.-Y., Zhang, Y.-Q., Liu, J., Wang, M., Liu, X., & Zhou, H. (2026). CUDA Agent: Large-Scale Agentic RL for High-Performance CUDA Kernel Generation.
• 2. Zhang, Z., Wang, R., Li, S., Luo, Y., Hong, M., & Ding, C. (2025). CudaForge: An Agent Framework with Hardware Feedback for CUDA Kernel Optimization. arXiv.org.